DeepMind paper co-evolves agent harness and weights, lifting SWE-bench 41.0% to 48.3%
omarsar0 · x · 2026-10-13
A new Google DeepMind paper proposes co-evolving an agent's harness with its weights, via a framework called Evo2:
- One policy plays two roles: a Proposer that writes edits to its own harness, and a Reflector that diagnoses what each edit changed
- Those edits and outcomes become training data for both roles, trained with SFT followed by interleaved GRPO
- Results: Evo2 hits 48.3% on SWE-bench Verified and 82.1% on Polyglot, vs 41.0% and 62.6% for the strongest baselines
- The harness alone moves a GPT-5-mini agent from 24.4% to 48.3% on SWE-bench Verified
Practical takeaway: if you already log your harness changes and results, those logs can become training data for the model running inside the harness.
Related event: DeepMind Paper Evolves Agent Harness Together with Weights(2 posts)→
More from coding & agent
- Isolating Antigravity chats per subfolder with AGY Rules and PreInvocation Hooks — bibryam · 2026-10-16
- Gemini shown off modding Minecraft to turn players into any Spider-Man — zacharynado · 2026-10-16
- Workflow: annotate UI screenshots in tldraw, hand to Claude, get better UI — max__drake · 2026-10-16
- Beyond chat wrappers: designing stateful, multimodal AI apps as interaction conductors — dr_cintas · 2026-10-16
- IBM engineer maps 5 levels of AI coding, from manual to a 3-day agent software factory — TejasKumar_ · 2026-10-16
- 10 open-source repos that let your AI agent reverse-engineer apps and games — victor_explore · 2026-10-16